BBoxStrategy Class

ml_debugger.data_filter.config.bbox_strategy.BBoxStrategy

Bases: BaseModel

BBox-level aggregation strategy for image-level error scoring.

属性：

名前	タイプ	デスクリプション
target_column	str	Error column to aggregate. One of 'error_proba' (default), 'det_error_proba', 'cls_error_proba'.
top_n	Optional[int]	Number of top-error bboxes per image. None (default) means no top-N restriction.
bbox_filter	Optional[Union[FilterConfig, Dict[str, Any]]]	Filter conditions for bbox selection. None (default) means no threshold filtering. If both bbox_filter and top_n are set, threshold -> top_n の順で適用。
aggregation	str	How to aggregate per-bbox values to image-level. 'mean' (default) or 'median'.
weight_column	Optional[str]	Optional weight column (e.g. 'pred_size'). - None: unweighted aggregation - set: weighted_mean / weighted_median に分岐
sort	str	Sort order for image-level results. 'desc' (default) or 'asc'.
img_error_threshold	Optional[float]	Image-level error threshold for real-time filter flags. None (default) means filter flags are all None (no filtering). If set (0.0-1.0), aggregated error_proba >= threshold → True, else False.

Example

Top 3 bboxes by det_error_proba, aggregated with max

strategy = BBoxStrategy( ... target_column="det_error_proba", ... top_n=3, ... aggregation="mean", ... )

Threshold + top_n + weighted aggregation by bbox size

strategy = BBoxStrategy( ... bbox_filter={"conditions": [{"error_proba": ">=0.3"}]}, ... top_n=3, ... aggregation="median", ... weight_column="pred_size", ... )

validate_dict(data) classmethod

Validate dict and return result with guide message.